str_rna_types = np.array([str(r) for r in rna_type])
# TODO is this what they meant?
def get_knn_score(input, gt_labels=str_rna_types):
max_accuracy = 0
for neighbor in range(1, 30):
knn = KNeighborsClassifier(n_neighbors=neighbor)
X_train, X_test, y_train, y_test = train_test_split(
input, gt_labels, random_state=42
)
knn.fit(X_train, y_train)
prediction = knn.predict(X_test)
accuracy = accuracy_score(y_test, prediction)
if accuracy > max_accuracy:
max_accuracy = accuracy
best_neighbor = neighbor
return best_neighbor, max_accuracy
# Note that the recall of the knn classifier is the same as the accuracy, for any of the averaging mode, as recall is a binary metric
return [
accuracy_score(y_test, prediction),
recall_score(y_test, prediction, average="micro"),
]
def get_all_knn_scores(result_dict, title, f=lambda x: x):
out_dict = dict()
for norm_key, sub_dict in result_dict.items():
out_dict[norm_key] = dict()
for trans_key, pca_result in sub_dict.items():
_, out_dict[norm_key][trans_key] = get_knn_score(f(pca_result))
return out_dict
print(
"Note that the recall of the knn classifier is the same as the accuracy, as recall is a binary metric\n"
)
high_dim_knn_score = get_all_knn_scores(pca_results, "PCA")
pca_knn_score = get_all_knn_scores(pca_results, "PCA 2D", lambda x: x[:, :2])
tsne_knn_score = get_all_knn_scores(tsne_results, "t-SNE")
umap_knn_score = get_all_knn_scores(umap_results, "UMAP")
knn_scores = [pca_knn_score, tsne_knn_score, umap_knn_score]
plot_score_table(knn_scores, "KNN Accuracy Scores")
